This EASE 2026 experience report examines how conversational “vibe coding” behaves when building two production-oriented AI systems under explicit architectural constraints: a multi-project agent learning platform with structured memory and strict tenant isolation, and a file-aware academic retrieval-augmented generation (RAG) system with citation tracing and role-based access control. Both systems were developed by a small full-stack team at Tampere University, with requirements — including tenant isolation, asynchronous execution, and controlled retrieval — defined before implementation began, and validated afterward through structured manual testing, code inspection, and runtime log analysis.
Across both systems, vibe coding accelerated scaffolding and feature implementation — API routes, serialization utilities, and interface components — but the generated code consistently under-specified architectural properties that were not explicitly stated in prompts: early routes omitted project-level filtering in database queries (risking cross-project data access), and memory-update and embedding routines initially executed synchronously despite an asynchronous requirement, both of which required manual correction. The paper reports a resulting shift in engineering effort away from boilerplate implementation and toward constraint specification, isolation auditing, and enforcement validation, and identifies recurring architectural areas — which it terms “non-delegation zones” — where conversational code generation remained insufficient for production reliability.
Key Points
- Reports that vibe coding reliably accelerated routine scaffolding (API routes, database models, UI components) in both systems, but rarely preserved architectural properties — tenant isolation, access control, and asynchronous processing — unless those constraints were made explicit in prompts
- Identifies a recurring shift in engineering effort: reduced effort on boilerplate writing, CRUD scaffolding, basic routing, and UI templates, and increased effort on architecture design, isolation auditing, policy specification, and validation/monitoring
- Proposes the concept of architectural “non-delegation zones” — recurring areas (multi-tenancy, access control, memory policies, asynchronous processing) that the authors found conversational code generation could not reliably handle without deliberate manual architectural design and verification
- Reports that infrastructure decisions such as background task queuing could not be safely deferred: early prototypes that ran inference and embedding synchronously became unstable under repeated use and had to be moved to background workers
- Argues, based on these two cases, that AI-assisted development is most effective as a tool for rapid implementation within clearly specified architectural boundaries rather than as a replacement for system-level design
Notes
- This is a retrospective, two-case experience report by a small team with prior full-stack and infrastructure experience; the authors state the findings are experience-based insights rather than broadly generalizable conclusions, and note their own involvement in building both systems as a potential source of bias, which they sought to mitigate by grounding the analysis in observable artifacts (commit history, runtime logs, documented prompts) rather than recollection
- No controlled experiments or quantitative productivity/code-quality/defect-rate measurements were conducted